Clinicopathological Machine Learning Models for Mortality Risk Prediction in Head and Neck Squamous Cell Carcinoma: A Retrospective Analysis of the The Cancer Genome Atlas Head-Neck Squamous Cell Carcinoma (TCGA-HNSC) Cohort.

BACKGROUND

Head and neck squamous cell carcinoma (HNSCC) is a leading contributor to global cancer mortality, yet outcome prediction in routine practice still relies almost entirely on anatomical staging. We evaluated whether readily available clinicopathological variables, and their combination with tumour transcriptomic data, can support machine learning (ML)-based mortality risk stratification in HNSCC.

METHODS

We analysed clinical records for 528 patients from the Cancer Genome Atlas Head-Neck Squamous Cell Carcinoma (TCGA-HNSC) cohort, obtained from the National Cancer Institute's (NCI's) Genomic Data Commons (GDC). Ten clinicopathological features were used to train four classifiers: logistic regression, random forest, gradient boosting, and support vector machine to predict vital status, evaluated by stratified five-fold cross-validation. Time-to-event outcomes were separately modelled with Kaplan-Meier estimation and multivariable Cox proportional hazards regression. In a matched subset of 500 patients with available tumour RNA-sequencing data, we additionally incorporated expression of four HNSCC driver genes (CDKN2A, EGFR, FAT1, and E2F1) identified from the TCGA landmark genomic analysis, together with the first five principal components of a 4,465-gene expression panel, and re-evaluated the same four classifiers on this matched cohort with and without the genomic features.

RESULTS

The cohort included 224 deaths (42.4%). Using clinicopathological variables alone, random forest achieved the highest discrimination (area under the curve (AUC) 0.698), followed by logistic regression (AUC 0.689), support vector machine (AUC 0.686), and gradient boosting (AUC 0.666). Kaplan-Meier analysis showed clear separation of overall survival by stage (log-rank Stage I vs. IV, p = 0.005), and multivariable Cox regression confirmed Stage IV disease (HR 4.07, 95% CI 1.50-11.03, p = 0.006) and older age (HR 1.02 per year, p = 0.005) as independent predictors of mortality (concordance index 0.608). Adding genomic features in the matched 500-patient cohort improved discrimination for every classifier, most notably random forest (AUC 0.708 → 0.727) and logistic regression (AUC 0.688 → 0.704); expression-derived features accounted for eight of the 15 most important predictors in the augmented model.

CONCLUSIONS

Routinely collected clinicopathological data provide a modest but genuine and statistically robust prognostic signal in HNSCC (AUC 0.67-0.70). Supplementing this with a small panel of driver-gene expression values and an unsupervised transcriptomic signature further improves discrimination (best AUC 0.727) without resolving the underlying prognostic ceiling, supporting stepwise, transparently benchmarked multi-omic model development rather than reliance on any single data modality.